Menu Menu

How to spot AI doctoring and deepfakes

At this point we’re all aware that AI doctoring and deepfakes are everywhere. Confidently identifying instances of their use, however, isn’t always straightforward.

AI has its mitts all over social media and the days of taking content immediately at face value are dying out.

There are obvious and largely harmless examples of AI manipulation – mostly used for meme purposes – but occasionally we find ourselves suspiciously squinting at an image or video to confirm its legitimacy.

The majority of us aren’t likely to be duped by a poorly rendered Mikel Arteta imposed over Lil Yachty bouncing around on stage, but there are technologies capable of creating awe-inspiringly realistic media.

In just the last few weeks, AI generated Met Gala images of Katy Perry and Rihanna went viral leading many to believe both artists attended the event – neither did. Elsewhere, an imitation audio program was used to mimic WPPs Mark Read and scam the company’s workforce.

The breakneck rate at which the technology is improving is quicker than we’re able to effectively regulate, and barring the odd clarification from community notes on Twitter, we’re basically left to decide for ourselves whether something is real or fake.

If you’re experiencing a sudden pang of paranoia, allow us to provide some helpful tips of what to look out for when examining content for evidence of AI. Soon enough, we’ll have to pointing at your screen like that Leonardo DiCaprio gif whenever AI rears its head.

 


Tips for detecting deepfakes

 

  • Look for obvious physical irregularities:

While AI is constantly tuning itself to perform better, the vast majority of existing deepfakes will have some irregularities if you know what you’re looking for.

Tell-tale signs include unnatural eye movement, weird blinking patterns or no blinking at all, and blurred or pixelated areas around the mouth, eyes, and hair.

Look for warping around the edges of the face, especially when the head turns, and if the quality of the video is high a fake will likely miss things like stray hairs or skin imperfections.

Often times, a deepfake will feature clunky and mechanical movements and you may even spot instances of a hand clipping through a sleeve or abnormal movement by the collar and neck.

 

  • Dodgy background and lighting:

Check if the lighting direction is consistent throughout the video. Artificially doctored examples often have weird patches of shadow and light that don’t necessarily correlate with light sources or objects in the background.

When the subject moves, keep a keen eye on the background to see if any distortions appear. If possible, go frame by frame and zoom.

 

  • Audio-visual mismatch

Consider the environment presented and whether or not the subject’s speech matches the acoustics of the space. The audio quality should also fit appropriately with the video resolution if legitimate. Both are big giveaway for your average deepfake.

It sounds obvious, but keep a close eye on the person’s lip movements. Deepfakes struggle to attain complete synchronisation and you’re sure to spot a notable blip if the video or clip is beyond a few seconds long.

 

  • Verify information and use AI to your advantage

Context is always key. Check online to see if reputable sources have ratified what you’ve seen on social media. If the video features a public figure doing something newsworthy, shock, it’ll show up when you search for news.

To be safe, you could always screenshot and use Google lens to see if the video has appeared anywhere else on the internet.

Lastly, there are many AI detection platforms such as Microsoft’s Video Authenticator, Deepware Scanner, or DeepFaceLab which may unearth things you didn’t notice.


Tips for detecting AI-doctored images

 

  • Uncanny valley

Is there something seemingly off about the people or objects within the frame? Poorly executed generations may have dodgy hands and fingers, and human skin often appears overly smooth. Teeth may lack separation completely.

Hair, fur, or other fine details can look unnaturally clumpy or blurry in AI-doctored images. Midjourney creations, for example, tend to have a cross-hatching effect within texture-dense parts of an image. Once you see it, you notice it everywhere.

Look at the edges of objects or people. Blurry, uneven, or pixelated edges usually signifies AI meddling.  Remember the missing segment of a sleeve on Kate Middleton’s eerie family photo?

Random – and sometimes non-existent – objects tend to crop up in AI generations too. If, say, a bedroom has been generated, there will likely be odd and warped looking objects on a desk or shelf.

 

  • Symmetry issues, repeating patterns, and reflections

When trying to fulfil prompts of an ‘ultra-realistic’ nature, we’ve noticed that AI struggles with creating realistic facial symmetry. It just seems to love overly perfect, defined features – for the people in focus, anyway.

For those in the background, features will likely be blurry. If there is a crowd, check for facial expressions that don’t match the context of the image, too.

In densely packed images you may twig repeating patterns. If you’re looking at a city, many of the windows, doors, or other architectural elements will likely be nigh-on identical. It’s part and parcel of generating content for prompts quickly.

This will apply to clouds, bricks, stones, vehicles, and just about anything that appears in multiples.

Some AI techniques generate images by stitching together patches from other content. Imperfect blending can often leave behind visible seams. Get the zoom tool out to find them.

Finally, AI really struggles to get reflections right. Check water within an image to see if there’s something slightly off about the reflection, and look at a subject’s eyes to see whether a reflection is present at all. If not, it’s probably fake. Sunglasses can be a dead giveaway.

 

  • Visual inspection tools

If you want to quickly check the credibility of an image, the reverse image search on Google is your best friend. If it doesn’t appear anywhere beyond maybe social media, it’s probably fake.

Better yet, pop the image into an AI detection tool. Our personal go-to is Deepware Scanner. Forensically and FotoForensics are also good at showing specifically where repeated patterns occur when the naked eye isn’t up to scratch.

That’s it from us, folks. We hope you’re now a little better equipped to navigate the ceaseless storm of online content without being caught out or misinformed. That being said, who knows if any of these tips will suffice five years from now?

Accessibility